Sentiment Analysis and Topic Modeling on MyGov Multiple Schemes Based Reviews

Topic modeling is widely used approach for identifying latent topics in large textual dataset. For improved topic extraction and representations, this work introduces an advanced topic modeling framework using BERTopic, which combines Large Language models (LLMs) such as Large Language Models Meta AI (Llama) and Mistral model, for dimensionality reduction and clustering, we have utilized Uniform Manifold Approximation and Projection (UMAP), and Hierarchical Density-Based Spatial Clustering of Applications with Noise (HDBSCAN) respectively. The proposed methodology starts with a Sentence-Transformer model to compute dense vector embeddings, then uses UMAP to reduce dimensionality and HDBSCAN for clustering. We have used Llama for topic labeling in order to enhance topic interpretability and guarantee meaningful representations of topics that were found. Coherence and perplexity metrices are used to assess our model’s performance, and we have compared the results across our MyGov reviews dataset, 20 Newsgroups, and the CAVES datasets. Furthermore, the performance of both our proposed LLMs-based framework and Latent Dirichlet Allocation (LDA) based framework for topic modeling have also been presented. In comparison to conventional methodologies, our results show that Llama with BERTopic generates greater coherence scores of 0.8407 at 3 topics, showing increased topic consistency and interpretability.